My journey

Hereford

Pool Cottage

Photo of Weobley High School / IMD

Bristol

Salamanca

York

Sheffield

Leeds

Sydney

Chapeltown

Chaco

Family

Family is the foundation on which many good things build.

Work

What is Transport Data Science?

(I-)Reproducibility

Reproducibility is a continuous variable (Peng 2011)

Why make your research (more) reproducible?

Why not make your research reproducible?

Example of fully reproducible research

Lovelace, Tennekes, and Carlino (2022)

Reproducibility and generalisability

Illustration of the ClockBoard zoning system used to visualize a geographically dependendent phenomena: air quality, measured in mass of PM10 particles, measured in micrograms per cubic meter, from the London Atmospheric Emissions Inventory (LAEI). The facets show the data in spatial grid available from the LAEI, facet Am and aggregated to London boroughs B, to ClockBoard zones covering all the input data shown in C, and ClockBoard zones clipped by the administrative boundary of Greater London in D.

Application: road traffic casualties

International comparisons

Case study: mobile telephone data in Spain

Don’t reinvent the wheel

options(timeout = 600) # 10 minutes
u1 = "https://movilidad-opendata.mitma.es/estudios_basicos/por-distritos/viajes/ficheros-diarios/2024-03/20240301_Viajes_distritos.csv.gz"
f1 = basename(u1)
if (!file.exists(f1)) {
  download.file(u1, f1)
}
drv = duckdb::duckdb("daily.duckdb")
con = DBI::dbConnect(drv)
od1 = duckdb::tbl_file(con, f1)

remotes::install_github("Robinlovelace/spanishoddata")
od_multi_list = get_od(date_regex = "2024030[1-7]")
# ...
n_per_hour |>
  ggplot(aes(x = Time, y = Trips)) +
  geom_line(aes(colour = Day)) +
  labs(title = "Number of trips per hour over 7 days")

Re-using existing tools

# Process the data
od_large = od_database |>
  group_by(origen, destino) |>
  summarise(Trips = sum(viajes), .groups = "drop") |>
  filter(Trips > 500) |>
  collect() |>
  arrange(desc(Trips))
# ℹ 37,013 more rows
# Convert to geo with {od} package:
od_large_interzonal_sf = od::od_to_sf(
  od_large_interzonal,
  z = distritos_wgs84
)
od_large_interzonal_sf |>
  ggplot() +
  geom_sf(aes(size = Trips), colour = "red") +
  theme_void()

Zooming in

distritos = get_zones(type = "distritos")
distritos_wgs84 = sf::st_transform(distritos, 4326)
salamanca_zones = zonebuilder::zb_zone("Salamanca")
distritos_salamanca = distritos_wgs84[salamanca_zones, ]
plot(distritos_salamanca)

Subsetting from the database

od_salamanca = od_database |>
  filter(origen %in% ids_salamanca) |>
  filter(destino %in% ids_salamanca) |>
  collect()
  group_by(origen, destino) |>
  summarise(Trips = sum(viajes)) |>
  arrange(Trips)
od_salamanca_sf = od::od_to_sf(
  od_salamanca,
  z = distritos_salamanca
)
od_salamanca_sf |>
  filter(origen != destino) |>
  ggplot() +
  geom_sf(aes(colour = Trips), size = 1) +
  scale_colour_viridis_c() +
  theme_void()

Spatial disaggregation

od_jittered = odjitter::jitter(
  od_salamanca_sf,
  zones = distritos_salamanca,
  subpoints = drive_net,
  disaggregation_threshold = 1000,
  disaggregation_key = "Trips"
)
od_jittered |>
  arrange(Trips) |>
  ggplot() +
  geom_sf(aes(colour = Trips), size = 1) +
  scale_colour_viridis_c() +
  geom_sf(data = drive_net_major, colour = "black") +
  theme_void()

Building/linking-up-with communities

The package has been onboarded to rOpenSpain public benefit data science community (see ropenspain.github.io)

How does spatial disaggregation (jittering) work?

Source: (Lovelace, Félix, and Carlino 2022)

Origin and end point randomisation + disaggregation

Source: (Lovelace, Félix, and Carlino 2022)

Cross-language collaboration

Source: https://github.com/dabreegster/odjitter

From open source to open access

From methods/software development to impact

Source: screenshot from development version of open source and open access Network Planning Tools for Scotland: https://nptscot.github.io/#/rnet/#9.29/55.9882/-3.4379

Rapid feedback loops and interactivity

Illustration of od2net client-side network generator (source: od2net.org)

Enabling participation

Source: acteng.github.io. Credit: Dustin Carlino (Alan Turing Institute and Active Travel England) and colleagues in ATE.

Drivers of demand for transport planning

transport planning software was originally designed in the late 1950s and onwards to plan for

increased use of cars [for personal travel], and trucks for deliveries and goods movement Boyce and Williams (2015)

Thankfully that is no longer a priority:

Policy drivers have changed dramatically since then: climate change mitigation, air quality improvement and public health are prioritised in the emergent ‘sustainable mobility paradigm’ Lovelace (2021)

How could/should/will demand shift in the future?

Stages of open and reproducible science

Conclusions

• Reproducible research is a key part of future-proofing transport planning, for your work, and for the discipline as a whole

• Open source software and open access tools are key to this, especially if you want to have humans in the loop

• AI is not a panacea, and has its own environmental costs

• In this context, there are some key desirable features of future-proof transport models and associated software and tools:

  • Open source
  • Open access
  • Reproducible
  • Human-in-the-loop
  • Easily adaptable to new data sources, methods, and demands

Future-proofing workflows Appendix

Avoid stranded assets (tech debt)

Source: Semieniuk et al. (2022)

Appendix: What if the machines do take over?

• Would you want the AIs to be trained on your work?
  • Initial thought: no way, that’s my data!
  • Second thought: if the AIs are going to take over, they might as well be well-informed, by good information and good intentions!
• Would you want to be able to understand how the AIs work?

AI takeover?

Where are we headed with AI?

Source: https://situational-awareness.ai

The limits of AI

Nezhurina (N.D.)

Environmental costs of ‘AI’

Source: theregister.com and Gupta et al. (2021)

Factoring-in build-out of AI data centres